Boltzmann nonuniqueness and exact local conservation

Source and verification

This article explains family 363 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

Overview

Family 363 studies whether entropy and physical conservation laws determine a unique evolution of a dilute gas. Its two manuscripts construct nonunique rough solutions; the later one adds exact local conservation of momentum and energy. Those stronger balances are a substantive addition to the earlier entropy-class result.

🏷️ Preliminaries: collisions and entropy

Let $F(t,x,v)\ge 0$ on ${\mathbb{T}}^3\times {\mathbb{R}}^3$. The hard-sphere equation is

$$({\partial }_t+v\cdot {\nabla }_x)F=Q(F,F).$$

For two velocities and a collision direction $\omega \in S^2$, set

$$\begin{array}{c}a=(v-v_{\ast })\cdot \omega ,\\ v^{\prime }=v-a\omega ,v_{\ast }^{\prime }=v_{\ast }+a\omega .\end{array}$$

The collision integral sums $|a|(F^{\prime }{F}_{\ast }^{\prime }-F{F}_{\ast })$ over $v_{\ast }$ and $\omega$, with all factors evaluated at the same position. Its positive and negative parts are called gain and loss.

For rough densities, finite mass, energy, and entropy need not make the quadratic collision integral integrable. A renormalized solution therefore first satisfies equations for functions such as $\beta (F/M)$, where $M$ is a Gaussian reference density and ${\beta }^{\prime }$ decays at infinity. This weak formulation must be specified when discussing uniqueness.

Relative entropy is the integral of $F\log (F/M)-F+M$. The entropy inequality bounds this quantity plus the accumulated nonnegative collision dissipation by its initial value. It is an admissibility condition, not by itself a uniqueness theorem.

🏷️ Two levels of nonuniqueness

The stronger claim in family 363

There exists a nonnegative initial density with bounded velocity support and finite mass, energy, and absolute entropy for which the periodic hard-sphere equation has two distinct global entropy solutions. Both are strongly continuous in $L^1$, satisfy exact local mass, momentum, and kinetic-energy conservation, and have separately integrable collision gains and losses with every polynomial velocity weight on finite time intervals. (OpenAI, 2026)

The earlier manuscript supplies the concentrating-jet nonuniqueness mechanism in an entropy class. Its global continuation does not supply all the later local balances. (OpenAI, 2026a) Bounded velocity support is an assumption on the initial density; it is not asserted to persist for the evolving gas.

The construction uses unbounded amplitudes on shrinking spatial scales. It therefore does not establish nonuniqueness for arbitrary smooth data or in the stronger classes where perturbative uniqueness arguments apply.

🏷️ Why collision integrability gives local conservation

The collision formulas directly imply

$$\begin{array}{rl}{v}^{\prime }+v_{\ast }^{\prime }& =v+v_{\ast },\\ |v^{\prime }{|}^2+|v_{\ast }^{\prime }{|}^2& =|v{|}^2+|v_{\ast }{|}^2.\end{array}$$

For the second identity, expand the squares: the change is $-2a(v-v_{\ast })\cdot \omega +2a^2=0$. Together with preservation of particle number, these identities make $1,v_1,v_2,v_3,|v{|}^2/2$ collision invariants.

When all integrations are justified, testing the equation against an invariant $\psi (v)$ gives

$${\partial }_t\int F\psi dv+{\nabla }_x\cdot \int vF\psi dv=0.$$

This is local conservation: it controls transport through every spatial region. A global energy inequality alone cannot replace it. The energy flux involves a third velocity moment, and passing the cancellation through a weak limit needs control of the separate gain and loss terms. This explains the role of the later manuscript’s weighted integrability estimate.

🏷️ Concentration, amplification, and global continuation

A cold jet at spatial scale $s$ can have velocity-integrated density of order $s^{-1}$ while its total mass is only of order $s^2$. A geometrically decreasing sequence of scales can therefore have finite total mass and entropy while producing large local collision frequencies. Alternating thin stripes of opposite streams converge under dilation to a two-stream background.

The proposed construction compares a dormant approximation with another carrying a seed whose initial mass tends to zero. The seed is tuned until an evolving norm reaches a prescribed bound. Rescaling at such a time must retain a nonzero limit; this is the function of the nonconcentration estimate. Branching growth in the limiting two-stream medium forces a finite rescaled horizon. Comparing that nonzero limit with the dormant solution’s vanishing limit separates the two solutions with identical initial traces.

The later paper must also continue both branches indefinitely. Periodic copies of a jet eventually meet, so each scale is treated as a separate component only for a finite lifetime. A nearly Maxwellian surrounding gas damps it afterward. Weighted estimates then transfer the decaying jet into a controlled remainder. Summability over scales yields integrable gains and losses, exact local conservation, and strong time continuity.

The decisive distinction is between a vanishing initial seed and a nonvanishing later rescaled response. Entropy bounds permit the concentration pattern; the detailed amplification and compactness arguments must prove that it survives the limit.

🏷️ Key steps and proof outlines

Dilation separates the linear amplifier from the nonlinear equation

The cold components are concentrated streams; the hot component is the perturbation they amplify. At a selected cap event, the construction rescales time and position by a small jet scale and normalizes the hot amplitude. In this regime the cold background converges to a measure $B$ concentrated near two opposite velocities. The hot self-interaction, bath error, seed source, and deviation of the cold cores from free transport are shown to vanish at the chosen normalization. The limiting equation is linear in the hot density $g$:

$$({\partial }_t+v\cdot {\nabla }_x)g=2P(B,g)-{\nu }_{B}g.$$

Here $P$ is the positive collision gain and ${\nu }_B$ its loss frequency. The factor two matters: a collision contributes both outgoing children while the loss removes the incoming hot particle once. The resulting growth is not obtained by dropping the nonlinear term without an estimate; each discarded term has its own bound in the dilation argument.

A cap hit must survive weak convergence

A normalized supremum reaching one does not ensure a nonzero weak limit: its support might shrink to zero measure. The nonconcentration argument addresses precisely this gap. Divide the rescaled hot density by its positive velocity majorant $H$ and write its positive gain inequality as a Volterra iteration. On a slab $a\le t\le 1$, the normalized gain kernel has mass at most $C/t$, giving

$$\Vert {\mathcal{V}}^{j}F(t){\Vert }_{\infty }\le \Vert F{\Vert }_{\infty }\frac{(C\log (t/a){)}^j}{j!}.$$

The factorial comes from ordering the collision times. It permits a controlled finite truncation of the iteration. If the weak limit were zero, positivity would give local $L^1$ convergence to zero. The manuscript then shows that the finite gain integral forced by the cap also tends to zero, a contradiction. Kernel tightness removes large velocities and Gaussian parameters; near the velocity axis, two gains supply the additional regularization that a single collision lacks. Away from the axis, finite-dimensional parameter changes supply uniform integrability. These are substantive collision-kernel lemmas, not consequences of the cap alone.

Branching growth beats the permitted space-time volume

The limiting gain has a useful probabilistic representation. In the reference two-beam medium, a collision sends $v$ to

$$\Phi (v,w,n)=\frac{v+w}{2}+\frac{|v-w|}{2}n,$$

with $n$ uniform on $S^2$ and $w=\pm e_3$. Restrict to velocities $|v|\le 5$. The source computes that a child remains in this ball with probability $p(v)\ge 25/26$. If ${\nu }_0$ is the reference loss rate, the retained branching histories equal a Markov-flight expectation with jump rate $2p{\nu }_0$ and multiplicative potential

$$\mathcal{W}=(2p-1){\nu }_0\ge 48\pi /13>0.$$

This identity follows by multiplying the Markov survival factor by $\exp (\int \mathcal{W})$: it becomes the original loss survival factor, while the jump weights remain the two-child gain weights.

Three suitable collisions give every starting velocity a common positive part of its transition law on a small ball. This mixing estimate controls displacement as well as velocity growth. Compact families of histories remain effective under the small measurable perturbations present in the actual medium. Iteration produces mass multiplication greater than three from $t$ to $(1+a)t$, with $(1+a{)}^4<3$. But the cap $g\le (t+l)H$ permits only order $t^4$ mass in a spatial ball of radius proportional to $t$: three powers from volume and one from amplitude. A nonzero capped limit therefore cannot persist for all rescaled times. This is the shared amplification mechanism of the two manuscripts. (OpenAI, 2026a; OpenAI, 2026b)

The later global extension uses a bath and retirement times

For the stronger result, write $F=c+w$, with $c$ the active cold jets and $w$ the bath, hot component, and retired jets. A jet at scale $s$ is retired at time $s^{-1/2}$ by transferring it from $c$ to $w$ without changing $F$.

Small entropy distance from the Maxwellian bath controls a spatial $L^1$ error, not a pointwise error. The bridge is free-flight averaging: for a bounded velocity set $V$ and $\tau \ge \rho >0$,

$${\int }_{V}A(x-\tau v)dv\le C_{V,\rho }\Vert A{\Vert }_{L_x^1}.$$

The change of variables contributes ${\tau }^{-3}$, while the number of periodic cells is $O((1+\tau {)}^3)$, so the constant remains bounded for large $\tau$. Duhamel’s formula then gives uniformly positive density in two separated velocity balls. Every velocity is separated from at least one of those balls, proving a loss frequency ${\nu }_F\ge \kappa >0$ independent of the bootstrap amplitude.

This damping makes retired jets summable and closes the bound on $w$. The resulting weighted uniform integrability of the separate gains and losses permits velocity averaging and passage through the collision-invariant cancellations. It supplies the later paper’s strong $L^1$ traces and exact local momentum and energy balances, which should not be retroactively attributed to the earlier theorem. (OpenAI, 2026b s “A uniform global bound for the finite evolutions,” “Global limits and local conservation,” and “Branching growth and the dilated horizon”)

See Also

• Family 362 concerns smooth collisionless dynamics and a different continuation criterion.
• Family 364 assumes a regular Boltzmann solution while deriving it from microscopic dynamics.

🏷️ Manuscripts and verification scope

• Nonuniqueness with local conservation for the hard-sphere Boltzmann equation — pinned manuscript and companion source.
• Nonuniqueness for the periodic hard-sphere Boltzmann equation — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The scope page selects the earlier periodic nonuniqueness manuscript. It covers local mass conservation, total momentum, and global energy and entropy inequalities, with strong continuity on a common initial interval. It explicitly excludes the later exact local momentum and energy conservation refinement and its stronger global continuity statement. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026a. Nonuniqueness for the periodic hard-sphere Boltzmann equation.

🐻  OpenAI 2026b. Nonuniqueness with local conservation for the hard-sphere Boltzmann equation.